Trust in AI output
Trust in AI-produced medical content turns on who is asked to rely on it, such as the regulatory reviewer who receives a submission prepared with AI, and on the conditions under which a system in use is trusted, such as who trained it and who corrects it.
Reviewers' trust in AI-assisted submissions
In May 2024, a speaker from the European Medicines Agency said it was not yet known whether regulatory reviewers trust submissions put together using AI more or less. He said that, short of the use of AI being obvious, which it would be only if it was done really badly, he was not quite sure, and that this was why it had been difficult to come to any real conclusion. He found it hard to say, but in his view the use of AI should not matter if a submission is accurate and its writers take responsibility for it, owning every line, full stop and comma even where AI generated them.
Conditions for trusting an AI assistant
In May 2024, a professor working on AI in oral healthcare described an AI dental assistant as fast, accurate, always on, unbiased and consistent: maybe not always correct, but with the clinician there at any moment to supervise it. She held that AI is often 10 to a thousand times faster for the dental practice than an assistant.
She held that, in her field, obtaining such an AI takes both a very highly trained engineer and a highly trained doctor. It needs supervised learning, in her account: the AI starts as a baby that is nursed and then goes to school, where the highest possible experts, rather than students or just traditional doctors, train it and put all their knowledge into it. She said the AI cannot get this knowledge from a black box except through big data, which are not supervised, and that her group uses only supervised learning because it is the only way the information can be medically trusted.
In her account, an AI trained for a long time can learn very efficiently from error. So, she said, once one starts working with AI, one also has to correct it at any moment, because correction is the only way it gets more and more performant. The correction has to come from the experts who are the clinical reference or clinical standard, not from a general practitioner.
As long as the AI starts from supervised learning by experts and stays supervised at all levels, with any error corrected, she held it to be a perfect dental assistant: fast, accurate, consistent and unbiased, not tired, on 24 hours, and always at the clinician's side at no cost.